From Accuracy to Interpretability in Autism Screening

Autism Spectrum Disorder (ASD) is a complex and increasingly prevalent neurodevelopmental condition that highlights the need for diagnostic tools that are both accurate and interpretable. While machine learning (ML) holds promise for ASD screening, limited transparency hinders clinical adoption. This thesis evaluates five ML classifiers—Logistic Regression (LR), Random Forest (RF), Extreme Gradient Boosting (XGBoost), TabNet, and Long-Term Cognitive Net- works (LTCNs)—on a synthetic ASD dataset under both imbalanced and Synthetic Minority Over-sampling Technique (SMOTE)-balanced conditions. Across models, standard metrics such as ROC-AUC, Weighted F1-score, precision-recall, and Cohen’s Kappa remained stable under balancing, with F1-scores between 0.85–0.91 and ROC- AUCs from 0.84–0.93. To assess interpretability, various feature attri- bution methods were compared. Native model importances capture internal feature rankings. Permutation Feature Importance (PFI) mea- sures performance drops when feature values are shuffled. Shapley Additive Explanations (SHAP) apply game theory to assign fair con- tributions to each feature. Performance Marginalization Contribution (PMC) quantifies accuracy changes when a feature is removed. Faith- fulness of explanations was evaluated with Performance Marginal- ization Curves (PMCs), where smaller area under the curve indicates closer alignment with model behavior. Additionally, Sparseness- Optimized Feature Importance (SOFI) was utlized which leverages PMCs to select a minimal yet relevant feature set, yielding interpreta- tions that are compact and consistent. By emphasizing interpretability, this work examines how ML-based ASD screening can lead toward tools that clinicians can understand, patients can trust, and families can rely on-supporting more transparent, and informed care.